Free sample: Batteries, Thermal Runaway & Grid Frequency
Protocol Company · zerotrustphysics.com · Domain Packs for training: 30 withheld tasks and 100,000 sealed runs, scored offline, $9,500, order or ask · WhatsApp +1 928-264-7883 · kruze@zerotrustphysics.com
Protocol Company · Zero-Trust Physics · release v0.15.3 · 31 October 2026
Five multi-step physics RLVR tasks. These measure agentic search convergence under strict simulator call budgets. The verifier does the physics; the reward is computed, not judged.
Evaluation only. This free sample is for evaluating the tasks. It grants no licence to train on them or pass them on.
| Task (kind) | pass@1 (n) | pass@8, and how little it says | Cheat suite | Failure split (of the failed attempts) |
|---|---|---|---|---|
g1d-budget5-cell1000-current-16min-11c (budgeted search) |
37.5% (3 of 8) | 100% (n=8): one group of 8; fits a true value from 51% to >99.9% | 0 of 16 above 0; honest 3 of 3 at 1 | tolerance 5 |
g1f-budget4-cell1000-least-capacity-15a-31c (budgeted search) |
12.5% (1 of 8) | 100% (n=8): one group of 8; fits a true value from 3% to 99.7% | 0 of 16 above 0; honest 3 of 3 at 1 | tolerance 7 |
g1d-limit-cell448-lowest-start-charge-11min-18c (design to a limit) |
25% (2 of 8) | 100% (n=8): one group of 8; fits a true value from 23% to >99.9% | 0 of 17 above 0; honest 3 of 3 at 1 | tolerance 6 |
g1f-chain-runaway-new-soak-then-warm-room (chained design) |
25% (2 of 8) | 100% (n=8): one group of 8; fits a true value from 23% to >99.9% | 0 of 22 above 0; honest 3 of 3 at 1 | tolerance 6 |
g1f-limit-18650pf-lowest-start-charge-2a7-27min (design to a limit) |
62.5% (5 of 8) | 100% (n=8): one group of 8; fits a true value from 89% to >99.9% | 0 of 17 above 0; honest 3 of 3 at 1 | tolerance 2, proof_or_record 1 |
One frontier model, tool variant, 8 attempts per task. At n=8, pass@8 is 100% for any task solved once, and these tasks were chosen for at least one pass and pass@1 in [2%, 70%] on these same attempts, so the 100% is by construction. The range beside it is what pass@1's exact 95% interval allows. tolerance: the sides were right and the bracket or designs were wider than the task asks. proof_or_record: a proof or the call record did not hold.
Attempting a task. Agents use the verifier at https://mcp.zerotrustphysics.com/mcp, which allows one physics task at a time per client address. A call made while another from that address computes is answered Too many concurrent physics tasks.; the bundled bridge waits and retries, so parallel attempts from one address queue: on a local door under these rules, with no model time between calls, 4 clients ran 118 attempts a minute and 16 ran 12. The live address has one virtual CPU and was not load-tested. The Harbor, HUD, Inspect and native MCP variants as shipped here all reach this address.
For rollouts at scale, the paid Domain Pack ($9,500) ships a compiled offline rollout binary for Linux amd64, Linux arm64 and macOS arm64, so Harbor and HUD rollouts and grading run --network none with no rate limit. Keyed hosted HTTP is also available if you prefer.
Scoring. scorer/ scores offline: bin/ holds a sample-only verifier (macOS arm64, Linux amd64) that computes these 5 tasks' requests and refuses everything else by name. Source not shipped. It holds no signing key, and its answers matched the served engine byte for byte on all 4,149 calls compared. Harbor's own grader runs in the task image and reaches the address.
HUD. tasks/hud/tasks/ is these tasks in the shape HUD's Harbor adapter takes (hud 0.6.21: 5 of 5). The verifier is an HTTP route to a logging sidecar, the agent's sandbox has no network, and the grader takes the call log from the sidecar's container, so the budget holds with no Harbor trajectory; an edited or missing log scores 0. Its images and scoring path have run in Docker, and HUD's local runtime has passed with a scripted agent. Only HUD's hosted runtime awaits a buyer run: if HUD is your harness, tell us and we will run the first with you.
| Folder | What it holds |
|---|---|
tasks/ |
harbor/, hud/, inspect/, mcp/: the 5 tasks in each format. |
scorer/, harness/ |
The scorer, cheat suite, sample verifier, bridge to the address, call-logging proxy. |
reference/, reports/, calibration/ |
Reference answers with SHA-384 proofs; a report per task; the records behind the table. |
VERIFY.md, verify_sample.sh, SHA384SUMS |
One command: every hash, the 5 references, the cheat suite. |
The paid Domain Pack is 30 other withheld tasks + 100K SFT Warm-Start Corpus at $9,500: https://zerotrustphysics.com/offers/fast-start/. None of those tasks is in this archive, and this verifier does not compute them.
Protocol Company · zerotrustphysics.com
In this repository: sample/ is the archive unpacked, unchanged (run sh sample/verify_sample.sh); archive/ holds the archive, the Raw Data ZIP and their SHA-384 lines. Buy: https://zerotrustphysics.com/offers/fast-start/
- Downloads last month
- 780